def lattice_to_dfa(lattice: pynini.Fst,
optimal_only: bool,
state_multiplier: int = 4) -> pynini.Fst:
"""Constructs a (possibly pruned) weighted DFA of output strings.
Given an epsilon-free lattice of output strings (such as produced by
rewrite_lattice), attempts to determinize it, pruning non-optimal paths if
optimal_only is true. This is valid only in a semiring with the path property.
To prevent unexpected blowup during determinization, a state threshold is
also used and a warning is logged if this exact threshold is reached. The
threshold is a multiplier of the size of input lattice (by default, 4), plus
a small constant factor. This is intended by a sensible default and is not an
inherently meaningful value in and of itself.
Args:
lattice: Epsilon-free non-deterministic finite acceptor.
optimal_only: Should we only preserve optimal paths?
state_multiplier: Max ratio for the number of states in the DFA lattice to
the NFA lattice; if exceeded, a warning is logged.
Returns:
Epsilon-free deterministic finite acceptor.
"""
weight_type = lattice.weight_type()
weight_threshold = (
pynini.Weight.one(weight_type)
if optimal_only else pynini.Weight.zero(weight_type))
state_threshold = 256 + state_multiplier * lattice.num_states()
lattice = pynini.determinize(
lattice, nstate=state_threshold, weight=weight_threshold)
if lattice.num_states() == state_threshold:
logging.warning("Unexpected hit state threshold; consider a higher value "
"for state_multiplier")
return lattice
def threshold_lattice_to_dfa(lattice: pynini.Fst,
threshold: float = 1.0,
state_multiplier: int = 2) -> pynini.Fst:
"""Constructs a (possibly pruned) weighted DFA of output strings.
Given an epsilon-free lattice of output strings (such as produced by
rewrite_lattice), attempts to determinize it, pruning non-optimal paths if
optimal_only is true. This is valid only in a semiring with the path property.
To prevent unexpected blowup during determinization, a state threshold is
also used and a warning is logged if this exact threshold is reached. The
threshold is a multiplier of the size of input lattice (by default, 4), plus
a small constant factor. This is intended by a sensible default and is not an
inherently meaningful value in and of itself.
Parameters
----------
lattice: :class:`~pynini.Fst`
Epsilon-free non-deterministic finite acceptor.
threshold: float
Threshold for weights, 1.0 is optimal only, 0 is for all paths, greater than 1
prunes the lattice to include paths with costs less than the optimal path's score times the threshold
state_multiplier: int
Max ratio for the number of states in the DFA lattice to the NFA lattice; if exceeded, a warning is logged.
Returns
-------
:class:`~pynini.Fst`
Epsilon-free deterministic finite acceptor.
"""
weight_type = lattice.weight_type()
weight_threshold = pynini.Weight(weight_type, threshold)
state_threshold = 256 + state_multiplier * lattice.num_states()
lattice = pynini.determinize(lattice,
nstate=state_threshold,
weight=weight_threshold)
return lattice
def all_suffixes(self, fsa: pynini.Fst) -> pynini.Fst:
fsa = fsa.copy()
start_state = fsa.start()
for s in fsa.states():
fsa.add_arc(
start_state,
pynini.Arc(0, 0, pynini.Weight.one(fsa.weight_type()), s))
return fsa.optimize()